Numpy array of numpy arrays has 1D shape

Question

I have two numpy arrays of arrays (A and B). They look something like this when printed:

A:

[array([0, 0, 0]) array([0, 0, 0]) array([1, 0, 0]) array([0, 0, 0])
 array([0, 0, 0]) array([0, 0, 0]) array([0, 0, 0]) array([0, 0, 0])
 array([0, 0, 0]) array([0, 0, 0]) array([0, 0, 1]) array([0, 0, 0])
 array([1, 0, 0]) array([0, 0, 1]) array([0, 0, 0]) array([0, 0, 0])
 array([0, 0, 0]) array([1, 0, 0]) array([0, 0, 1]) array([0, 0, 0])]

B:

[[  4.302135e-01   4.320091e-01   4.302135e-01   4.302135e-01
    1.172584e+08]
 [  4.097128e-01   4.097128e-01   4.077675e-01   4.077675e-01
    4.397120e+07]
 [  3.796353e-01   3.796353e-01   3.778396e-01   3.778396e-01
    2.643200e+07]
 [  3.871173e-01   3.890626e-01   3.871173e-01   3.871173e-01
    2.161040e+07]
 [  3.984899e-01   4.002856e-01   3.984899e-01   3.984899e-01
    1.836240e+07]
 [  4.227315e-01   4.246768e-01   4.227315e-01   4.227315e-01
    1.215760e+07]
 [  4.433817e-01   4.451774e-01   4.433817e-01   4.433817e-01
    9.340800e+06]
 [  4.620867e-01   4.638823e-01   4.620867e-01   4.620867e-01
    1.173760e+07]]

type(A), type(A[0]), type(B), type(B[0]) are all <class 'numpy.ndarray'>.

However, A.shape is (20,), while B.shape is (8, 5).

Question 1: Why is A.shape one-dimensional, and how do I make it two-dimensional like B.shape? They're both arrays of arrays, right?

Question 2, possibly related to Q1: Why does printing A show the calls of array(), while printing B doesn't, and why do the elements of the subarrays of B not have commas in-between them?

Thanks in advance.

Answer 1

A.dtype is O, object, B.dtype is float.

A is a 1d array that contains objects, which happen to be arrays. They could just as well be lists or None`.

B is a 2d array of floats. Indexing one row of B gives a 1d array.

So A[0] and B[0] can appear to produce the same thing, but the selection process is different.

Try np.concatenate(A), or np.vstack(A). Both of these then treat A as a list of arrays, and join them either in 1 or 2d.

Converting object arrays to regular comes up quite often.

Converting a 3D List to a 3D NumPy array is a little more general that what you need, but gives a lot of useful information.

also

Convert a numpy array of lists to a numpy array

==================

In [28]: A=np.empty((5,),object)
In [31]: A
Out[31]: array([None, None, None, None, None], dtype=object)
In [32]: for i in range(5):A[i]=np.zeros((3,),int)
In [33]: A
Out[33]: 
array([array([0, 0, 0]), array([0, 0, 0]), array([0, 0, 0]),
       array([0, 0, 0]), array([0, 0, 0])], dtype=object)
In [34]: print(A)
[array([0, 0, 0]) array([0, 0, 0]) array([0, 0, 0]) array([0, 0, 0])
 array([0, 0, 0])]
In [35]: np.vstack(A)
Out[35]: 
array([[0, 0, 0],
       [0, 0, 0],
       [0, 0, 0],
       [0, 0, 0],
       [0, 0, 0]])

Edit

np.stack(A)

can join the arrays on a new leading axis.

If the subarrays differ in shape, these 'stack' functions will raise an error. It's up to you to find the problem array(s).